[screenleon]

Refs #783 (CC-696). Fixes the three memory-usage/sqlite-* cases. The two concurrency cases in the report did not reproduce here, so #783 stays open for them (the #738 acceptance of 50 parallel writers is still unverified on native Windows).

根因

資料庫其實寫對了（手動重現：card_usage 有 a.md|1|100，metadata 有 total_events）。是測試讀它的方式錯了：

1. $($real_sqlite "$db" ...) 沒有加引號。real_sqlite 是 /c/Users/Lien Chen/AppData/.../sqlite3，路徑含空白就被拆開，指令找不到、輸出為空，報告成 count=missing。
2. 原生 sqlite3 輸出 \r\n，bash 的命令替換只去掉結尾的 \n，指令成功也會多一個 \r，與預期字串不相等。

修法

三個案例的 $real_sqlite 加引號，輸出接 tr -d '\r'。產品程式 memory_usage_commit 沒有改。BACKLOG 加 CC-696，並把 CC-695 記上 pr:#786。

驗證（原生 Windows，使用者目錄 Lien Chen 含空白）

• test-guards.sh --filter memory-usage：main 上 12 通過 3 失敗（三個 sqlite 案例），改後 15 通過 0 失敗。
• 突變就是 main 本身：還原這個檔案，三個案例失敗。
• 兩個並行案例（issue 說 8 個 writer 逾時 20 秒）在 main 上這台機器就通過：concurrent-no-lost-updates 約 12 秒、contention-matrix 約 34 到 46 秒，連跑三次都過。我沒有改它們，也不能說它們已修；它們的 20 秒期限在負載高時可能不夠。
• 沒有跑 pr-gate（本機規則：未經要求不跑）。

Review (commands/ship.md Step 2.5)

PASS: no warranted refactor/reuse change: the three sites are one-line quote and pipe changes in test code; a shared query helper would be a new abstraction for three lines.

• Permanent test admissions: no pr-gate round was run, so there is no gate finding; no test function was added or removed. The three existing cases only changed how they read the database (quoting, CR), not what they assert.

🤖 Generated with Claude Code

Review (commands/ship.md Steps 2.5 and 3)

• Refactor/reuse audit: PASS: no warranted refactor/reuse change. No sqlite or CR helper exists in tests/lib; the three sites are one-line quote and pipe changes, a shared helper would be a new abstraction for three lines.
• Review executor/model: codex (gate-resolved), tier standard (auto), mode sequential, reviewers: critic, qa-tester, architecture-reviewer, security-reviewer (risk-reviewer was not selected; the report claims nothing on that dimension)
• Rounds: 1
• Final verdict: GO (critic approve, qa-tester pass, architecture approve, security pass; Must-Fix: none, no finding id, all four test-gap rows no_gap)
• Result file: gate-20261105-143957.md (gate-20261105-053918-d811d2), pmctl gate wait verified it
• Cautions kept as is: the review workspace had no git metadata (scope relied on the manifest and SHA256 snapshots); the pre-flight evidence covers the 15-case memory-usage group, not the whole test-guards suite; the 50-writer native-Windows acceptance limit stays open (CC-696 "not reproduced" note).